Information Technology Reference
In-Depth Information
where
S
k
is
a unitary
mode matrix
of dimensions
N
k
×
N
k
, spanning the column space
of
the
matrix
T
(k)
obtained
from
the
mode-n
flattening
of
T
;
……
Z
is
a core tensor
of the same dimensions as
T
, which satisfies
the following conditions [16]:
1. Two subtensors
NN N N N
××
×
×
∈ℜ
1
2
m
n
P
Z
and
Z
, obtained by fixing the
n
k
index to
a
, or
b
,
k
na
=
k
nb
=
are orthogonal, i.e.
ZZ
,
⋅
=
0
(2)
na nb
=
=
k
k
for all possible values of
k
for which
a
≠
b
.
2. All subtensors can be ordered according to their Frobenius norms
ZZ Z
,
≥
≥
…
≥
≥
0
(3)
n
=
1
n
=
2
n N
=
k
k
k
P
for all
k
.
The following Frobenius norm
k
na a
Z
=
σ
(4)
=
k
is called the
a-mode
singular value of
T
. Each
i-th
vector of the matrix
S
k
is the
i-th k-
mode
singular vector.
Assuming decomposition (1) of a tensor
T
, singular values (4) provide a notion of an
energy of this tensor in the terms of the Frobenius norm, as follows
R
R
()
( )
2
1
2
P
2
2
∑
∑
1
P
T
=
σ
=
=
σ
=
Z
,
(5)
…
a
a
a
a
=
1
=
1
where
R
k
denotes a
k-mode
rank of
T
.
The SVD decomposition allows representation of a matrix as a sum of rank one
matrices. The summation spans number of elements, however no more than a rank of
the decomposed matrix. Similarly to the SVD decomposition of matrices, based on
the decomposition (1), a tensor can be represented as the following sum
N
P
∑
s
T
=
T
h
hPP
×
,
(6)
h
=
1
where
TZ
=× ×
SS S
×
,
(7)
…
h
1122 1 1
P
−
P
−
denotes the basis tensors and
s
h
P
are columns of the unitary matrix
S
P
. Since
T
h
is of
dimension
P
-1 then
P
in (6) is an outer product, i.e. a product of two tensors of
dimensions
P
-1 and 1
.
The result is a tensor of dimension
P
, i.e. the same as of
T
.
Fig. 2 depicts a visualization of this decomposition for a 3D tensor. In this case
T
h
×
Search WWH ::
Custom Search